run yolov5

跑跑yolov5

If you just want to get started with training quickly and get training results,我的建议是直接使用b导的代码.别问为什么,Ask is I ambGuide the little brother！！
但是
bThe guide code seems to be written by myself,会存在：The downloaded weights file ispth,最后得到的文件是pth文件 的问题 In fact, it may just be that I can't convert myselfhhh

Basically it can be divided into four steps：环境配置,数据集制作,yolov5Document modification to train,预测.I mainly talk about data set production,yolov5文档修改.Because I stepped on more pitshhhh

（1）使用labelimg数据集制作

yolov5不接受xml文档,But the strange thing on the Internet is that most bloggers like it“曲线救yolov5”（先制作xml,在转化成txt）

直接改成yolov5支持的yoloThe format is not good！！！

I only had one class to train so I just opened these straight up！！！

①Auto SAVE Mode：If you don't turn it on, you have to press it every times保存,Otherwise, there will be a pop-up reminder
But doing this is risky,It may appear that the hand speed is too fast to jump one or two pictures（Don't ask how I know that）
②Single Class Mode：因为我只有一个类,You don't have to choose to use this

Later, if I remember, I will write a script that checks for less labelinghhhhh

③Display Label：Display the tag name
The advantage of this is that there is a high probability that repeated annotations can be avoided！！

（2）yolov5文档修改

Let's talk about the newly created folder structure first

修改yaml文件

因为我懒,So I didn't recreateyaml文件,我是直接在data/coco128.yamlmodified in the folder

1. 改路径（文件路径,训练集路径,val路径）
   直接pycharmRight-click on the selected foldercopy path–>absolute path,Absolute paths are not easy to report errors
2. 改nc（标签数量）,name（标签名字)

修改train文件

定位到parse_opt,Only talk about modified files（wrong again）

1. 改pt （Absolute paths are not easy to report errors）

 parser.add_argument('--weights', type=str, default=ROOT / 'D:/computervision/deep_learning_model/yolov5-6.1/yolov5s.pt', help='initial weights path')

2. –data

parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path')

3. –epochs
   The dataset I train on is relatively small,–epochsMust be kept small,不然容易过拟合

parser.add_argument('--epochs', type=int, default=100)

跑yolov5What went wrong again

Just a word of mouth：Really hate the red tips and error oneshai！！！

（1）错误报告1：TensorBoard logging requires TensorBoard version 1.15 or above

Actually my version I was over2.6,It is estimated that there is a problem with another dependency package

1. Check for download firsttensorboard
   Enter the following command in the environment,检查有没有tensorboard

conda list

2. If the version has been exceeded1.15,那就Go to this URL
   Or directly on the command line

pip install tb-nightly==2.8.0a20220120

Here I basically basicallyok了

（2）错误报告2:AssertionError: No images found in ×××××

This mistake is a bit silly,I put the label's folder in it“val"后面,Then it says no picture found

留个小问题：yoloHow to get the label file？

检查一下yolov5the training results

可能是因为wandbVersion incompatibility is also possible这个原因
My file will compare别家的少.

从下往上看：

1. Three training result graphs
2. result.csv
   The first row is from left to right①epoch,
   ②train/box_loss,③train/obj_loss,④train/cls_loss,
   ⑤metrics/precision,⑥metrics/recall,⑦metrics/mAP_0.5,⑧metrics/mAP_0.5:0.95,
   ⑨val/box_loss,⑩val/obj_loss, ①①val/cls_loss,
   ①②x/lr0,①③x/lr1,①④x/lr2
   

(1)train：训练集;val：验证集
(2) cls_loss用于监督类别分类,box_loss用于监督检测框的回归,obj_loss用于监督grid中是否存在物体.(You can take a look at what this blogger has to say)
(3)！！！！！主要看这个！！！！！
mAP_0.5 : IoU设为0.5时,计算每一类的所有图片的AP,然后所有类别求平均,即mAP
mAP_0.5:0.95 : 表示在不同IoU阈值（从0.5到0.95,步长0.05）（0.5、0.55、0.6、0.65、0.7、0.75、0.8、0.85、0.9、0.95）上的平均mAP.

1. opt.yaml
   There's not much to say about the config file
   
2. hyp.yaml
   Estimates are hyperparameters
   
3. There is also the weight file

看到一篇博文对于trainThe parameters are very good, so I want to record them
（1）“–image-weights”
是否启用加权图像策略,默认是不开启的;主要是为了解决样本不平衡问题;After it is turned on, it will be used for the previous round of training效果不好的图片,在下一轮中增加一些权重;

1.28"–label-smoothing"
Whether to smooth the labels,默认是不启用的;

在训练样本中,我们并不能保证所有sample都标注正确,如果某个样本标注错误,就可能产生负面印象,如果我们有办法“告诉”模型,样本的标签不一定正确,那么训练出来的模型对于少量的样本错误就会有“免疫力”采用随机化的标签作为训练数据时,损失函数有1-ε的概率与上面的式子相同,比如说告诉模型只有0.95概率是那个标签.